PERFORMANCE COMPARISON OF RIVER WATER LEVEL PREDICTION USING SPARSE MODELING ASSUMING HEAVY RAIN DISASTER

Abstract

In this paper, we aimed to use sparse modeling to predict river water levels during heavy rain disasters. To achieve this goal, we compared the Lasso regression, which considers only the L1 regularization term, the Ridge regression, which considers only the L2 regularization term, and the Elastic net, which generalizes these two regularization methods. The water level prediction model by each algorithm was generated by learning the data for 6 points in 3 rivers for the past 10 years. Then, we evaluated the performance of the model using the observed values of the heavy rain disaster on July 2nd year of Reiwa. As a result, the Lasso and Ridge regressions are weighted to different explanatory variables. However, Elastic net has almost the same predictive performance as Lasso regression, and many of the explanatory variables selected are common. When the basin area was large, the prediction using sparse modeling tended to have a long lead time. Furthermore, we compared sparse modeling and LSTM recurrent neural networks using the observations of the 2018 Hokkaido heavy rain disaster. As a result, we found that predictions with almost the same accuracy can be made.